If your team plans to run Anthropic’s agent platform on its own infrastructure, the license matters as much as the architecture. Self-hosting keeps credentials and session history in your environment, but a written safety notice from Anthropic can start an eight-hour clock to disable a feature or roll back across every deployment. Platform teams need an on-call path, a rollback process and legal sign-off before they pilot it.
What happened
Anthropic has launched self-hosted Claude Managed Agents in private beta, according to The New Stack. Customers run both the server that coordinates agents and the sandboxes where agents execute tools, in their own cloud account or an on-premises data center. Session histories can sit in the customer’s PostgreSQL database. Requests to Claude go to whichever inference provider the customer configures, including Anthropic’s API, Amazon Bedrock, Microsoft Foundry or a company gateway.
The New Stack reports that the license carries tight conditions. The server ships as a proprietary binary that customers cannot modify. Models from other developers are barred unless Anthropic permits them. If Anthropic flags a safety, security or misuse risk in writing, customers have eight hours to disable the affected feature or roll back the version everywhere, and one business day to install a designated safety update.
The license also lets Anthropic decide, on safety grounds, that some Claude models are reachable only through its hosted service. Customers cannot block or alter "usage signals", the observability data the software may send to Anthropic for usage measurement and compliance monitoring.
Why it matters
Managed Agents now comes in three forms: fully hosted, hosted with customer-run sandboxes, and fully self-hosted. The last step removes much of the platform-building work that some companies, such as Spotify with its Honk system, did themselves. Agents that repair builds or triage alerts need access to private repositories and release systems, so keeping the control plane in-house is a real draw.
The trade is control over change timing. Patrick Moorhead of Moor Insights & Strategy told The New Stack that self-hosting usually means the customer owns the change window, and that this license flips that. His colleague Jason Andersen called the terms a reasonable ask and said buyers needing full control would likely look at open-weight options anyway.
This suggests a split in the market. Teams already comfortable consuming Claude as a cloud service will treat this as that service placed in their data center. Banks, government and sovereign buyers, as Moorhead expects, are likelier to contest the rollback and telemetry clauses. The practical logic is clear: an agent running on your servers can keep acting after the model provider has fixed something on its side, so the customer has to contain it.
What it means for you
Start with your change process. Check whether you can roll back or disable an agent feature across every environment in eight hours, including nights and weekends. If your normal production change takes longer, plan for an emergency path and test it before the beta moves to real workloads.
- Name who receives written safety notices and make sure that mailbox is monitored around the clock.
- Inventory every deployment, since the deadline applies across all of them.
- Ask legal and security to review the safety rollback and usage-signals sections early, not at procurement’s end.
- Confirm your data policy allows usage telemetry to leave your network, since you cannot block it.
- Accept that you cannot patch the binary yourself, so fixes depend on Anthropic’s release timing.
Also weigh the model limits. If your strategy relies on mixing vendors, the Claude-only rule and Anthropic’s right to reserve some models for its hosted service reduce flexibility. Budget for that lock-in alongside infrastructure costs, and revisit how you measure AI spend per task when you compare hosted and self-run options. Teams weighing where agent workloads should run can also read our workload test for moving agents to EKS.
What to watch
- Whether Anthropic softens the eight-hour window or adds tiers for regulated customers when the beta widens.
- How banks, government and sovereign buyers respond on the rollback and usage-signal clauses.
- Whether rival model labs copy these license terms for their own self-hosted agent products.
- Any real safety notice issued under the clause, which would show how disruptive the deadline is in practice.
